Pieter Levels launched Avatar AI as an ‘AI mini project‘ in October, 2022. The service quickly went viral and surpassed the income he made from his other projects.
He had his first sale within 17 minutes of announcing the launch. An hour later, he had made 6 sales.
He has been building the product in public.
Types of photos
Pieter also added different types of problems in response to audience requests:
- Couple photos. Some problems with MM and FF couples cause AI likes to revert back to MF, will add selector and fix tomorrow
- Dogs and cats
- Style selector. People could create a package of the type of styles they wanted to generate. It was hosted on Typeform.
- 4K high-resolution avatars. He added a added 20% off link when launching the avatars. He later described the new renders as insanely detailed and photorealistic
- LinkedIn-style professional headshots you can use for work. You can see photos of Pieter as a business woman
- Added Halloween and Christmas in response to holiday demand
He made $10k within a day from over 139 customers. This includedover $4100 in first 17 hours.
These sales were mostly from Twitter. However, Instagram quickly became the top referrer after it went viral on the platform
After 10 days, he had made $100’000 USD from Avatar AI. He had made 2,943 sales with an average of $33 per sale. He had paid around $25k in GPU costs, which included $8k in just one day.
Pieter has done a couple of marketing experiments.
He launched Avatar AI on Product Hunt on November 2. It became the #2 Product of the Day with over 900 upvotes. View the launch.
He offered a special Halloween promotion. Customers could get %25 off.
He added a referral link. People could get extra avatars if they shared the linked on social media.
He shared a Twitter thread about the risks of train models leaking from AI apps.
He made a robot that auto deleted all photos, and the entire AI model, 24 hours after rendering.
Spent an evening making making a robot to process the unfinished jobs. Some people wouldn’t get their avatars as the GPU image library occasionally died due to bad input files. He made a mini dashboard to keep track of GPU jobs to try and reduce the failed jobs from 5% to %0.
He temporarily took the API that he used to train the models down after too many people signed up.
He quickly saw ads on Upwork from people wanting to build similar websites.